LGD: Label-Guided Self-Distillation for Object Detection

This is the official implementation of the AAAI 2022 paper [LGD: Label-Guided Self-Distillation for Object Detection] (https://arxiv.org/abs/2109.11496)

Introduction

TL; DR We propose the first self-distillation framework for general object detection, termed LGD (Label-Guided self-Distillation).

Abstract. In this paper, we propose the first self-distillation framework for general object detection, termed LGD (Label-Guided self-Distillation). Previous studies rely on a strong pretrained teacher to provide instructive knowledge that could be unavailable in real-world scenarios. Instead, we generate an instructive knowledge by inter-and-intra relation modeling among objects, requiring only student representations and regular labels. Concretely, our framework involves sparse label-appearance encoding, inter-object relation adaptation and intra-object knowledge mapping to obtain the instructive knowledge. They jointly form an implicit teacher at training phase, dynamically dependent on labels and evolving student representations. Modules in LGD are trained end-to-end with student detector and are discarded in inference. Experimentally, LGD obtains decent results on various detectors, datasets, and extensive tasks like instance segmentation. For example in MS-COCO dataset, LGD improves RetinaNet with ResNet-50 under 2x single-scale training from 36.2% to 39.0% mAP (+ 2.8%). It boosts much stronger detectors like FCOS with ResNeXt-101 DCN v2 under 2x multi-scale training from 46.1% to 47.9% (+ 1.8%). Compared with a classical teacher-based method FGFI, LGD not only performs better without requiring pretrained teacher but also reduces 51% training cost beyond inherent student learning.

Main Results (upon MS-COCO across regular and supplementary paper sections)

Experiments are mainly conducted with 8x 2080 ti GPUs. We provide results (Table 1, 12 and 13 in the arXiv version) of common detection heads with various backbones equipped with FPN. Particularly, for the usage of Swin-Tiny backbone originally experimented under mmdetection environment in detectron2, you may conduct a conversion of its ImageNet-pretrained weights. We have done it for you and the converted weight file is available at LINK. Simply create a $pretrained_backbones$ sub-directory under ${PROJ} and put the ".pth" file under it. We re-run the experiments after a basic code-refactoring for higher readability. The results are consistent, with only 0.1 mAP difference (+0.1 mostly) compared to that exhibited in the arXiv version. Accompanied with the codes and results, we also release the relevant pretrained models and logs below.

RetinaNet

FCOS

Backbone | mAP | config | log | pretrained model --- |:---:|:---:|:---:|:---:| R-50 | 42.4 | config | LINK | LINK | R-101 | 44.0 | config | LINK | LINK | R-101-DCN v2 | 46.3 | config | LINK | LINK | X-101-DCN v2 | 47.9 | config | LINK | LINK |

Faster R-CNN

Backbone | mAP | config | log | pretrained model --- |:---:|:---:|:---:|:---:| R-50 | 40.5 | config | LINK | LINK | R-101 | 42.2 | config | LINK | LINK | R-101-DCN v2 | 44.8 | config | LINK | LINK | X-101-DCN v2 | 46.2 | config | LINK | LINK |

Mask R-CNN

Backbone | mAP(box) | mAP(mask) | config | log | pretrained model --- |:---:|:---:|:---:|:---:|:---:| Swin-Tiny | 46.4 | 42.5 | config | LINK | LINK |

Installation

This codebase is built upon [detectron2] (https://github.com/facebookresearch/detectron2)

Requirements

Ubuntu 16.04 LTS, CUDA>=10.0, GCC>=5.4.0
Python>=3.6.12
Virtual environment via Anaconda (>=4.10.3) is recommended:
```
conda create -n lgd python=3.7
```
Activate it by
```
conda activate lgd
```
detectron2==0.3
Pytorch>=1.7.1, torchvision>=0.8.2
Other requirements
```
pip3 install -r requirements.txt
```
Get into the LGD code directory (denoted by ${PROJ}).
```
cd ${PROJ}
```

Usage

Dataset preparation

mkdir ${PROJ}/datasets
ln -s /path/to/MSCOCO datasets/coco

Training

Single Machine

python3 train.py --config-file ${CONFIG} --num-gpus ${NUM_GPUS} --resume

Notes: We normally use 8 gpus at once for each experiment and that means ${NUM_GPUS}=8.

Multiple Machine

For experiments with backbone Swin-Tiny that are proned to be OOM, we opt for built-in distributed training with two machines supported by Pytorch or just a s

LGD

Install / Use

README